We stopped training LLMs from scratch. We built an AI Foundry instead.

Описание и применение

Малые команды не могут позволить себе обучение LLM с нуля, но им нужны безопасные и эффективные модели для конкретных отраслей. Компания VIDRAFT создала «AI Foundry», где диагностируют модели, комбинируют лучшие качества открытых моделей, проводят детальные проверки безопасности и оптимизируют для локального развертывания, что привело к высоким скачкам и победе в конкурсе Fast Gemma.

Задача

Small teams cannot afford to pretrain competitive LLMs, but they need domain‑specific models that are trustworthy and efficient.

Результат

The AI Foundry approach enabled an on‑device model with 1.18M cumulative downloads and a first‑place finish in the Google x Hugging Face Fast Gemma Challenge (510.58 tok/s).

Рабочий процесс

• 1. Gather domain requirements from the customer.
• 2. Select complementary open models with desired strengths.
• 3. Measure per‑capability deltas using behavioral, representation, and supply‑chain probes.
• 4. Recombine or distill models along axes where each parent excels.
• 5. Perform safety and reliability diagnostics (hallucination, jailbreak resistance, tool‑calling determinism, hidden‑state checks, serialization safety).
• 6. Quantize and optimize for on‑device or on‑prem deployment.
• 7. Validate the final model against benchmarks and real‑world usage metrics.
• 8. Deploy and monitor for drift, re‑heal if necessary.

Доказательства

An on‑device series built this way passed 1.18M cumulative downloads on Hugging Face, and our verified entry took first place in the Google x Hugging Face Fast Gemma Challenge (510.58 tok/s, verified track, 2026-08-02).

Источник

The row people skip is the fourth. Anyone can fine‑tune. Very few can tell you whether the resulting model is trustworthy, with numbers.

Источник